{"id":"717a07d4-cb4b-4f6a-b0f0-d716c1800beb","arxiv_id":"2504.19672","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"From 36 neutron-star X-ray binaries, the best-fitting natal kick model is a bimodal Maxwellian with sigma=320 km/s for core-collapse supernovae and sigma=80 km/s for electron-capture supernovae.","lead":"Using new and existing observations of 36 neutron-star X-ray binaries, this study reconstructs how fast these systems were moving when the neutron stars were born. The authors conclude that the best-fitting natal kick model combines fast kicks near 320 km/s and slower kicks near 80 km/s, a result that helps calibrate neutron star formation models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Best-fit claim rests on a 3D KDE likelihood that is -inf for most models; the ranking is not robust until bandwidth and uncertainties are tested.","rationale":"The reader's verdict is CONDITIONAL, and my read supports that conditionality rather than overturning it. The reader's weakest_assumption focuses on the vz=0_pec proxy for vsym; that is a real astrophysical concern. However, the more immediately load-bearing problem is internal to the paper's own statistics: the 3D likelihood used to select the headline model returns -inf for most models, which cannot be a legitimate Gaussian-KDE result and indicates numerical underflow. This means the central 'best model' claim is not supported by the evidence as presented, regardless of how the velocity proxy behaves. I partially agree with the reader because they also notice the -inf entries and describe the density estimate as flawed, but their stated weakest assumption is the velocity proxy rather than the likelihood defect. The paper has genuine strengths: a larger HMXB sample, a reasonable BPS setup, and an honest discussion of uncertainties and of the Swift J0243.6+6124 outlier. Those strengths justify a conditional acceptance with required statistical revision, not rejection. The concrete test is designed to settle whether the sigma2=80 ranking survives a proper likelihood calculation.","tokens_in":23406,"tokens_out":4948,"duration_ms":54719,"concrete_test":"Recompute Table 3's log Lambda with a leave-one-out cross-validated KDE bandwidth (and also with Silverman's rule and twice Scott's rule), and evaluate each observed point using random draws from its asymmetric 16-84% vz=0_pec uncertainties instead of the median, averaging log Lambda over the draws. Repeat with and without Swift J0243.6+6124. If sigma2=80 is not the highest finite likelihood in all variants, or if the -inf entries become finite and reorder, the claimed best-fit is a KDE artifact rather than a robust inference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 5 claim a specific best model (sigma1=320, sigma2=80, Mecs=1.83-2.25), but this is not established by the velocity KS test in Section 4.1, which the authors themselves say cannot directly select a best model. The selection instead comes from the 3D likelihood in Section 4.2, Eq. 11, evaluated with a Gaussian KDE using Scott's rule. Table 3 shows log Lambda(Porb,e,vz=0_pec) = -inf for 8 of 11 models in the Mecs=1.83-2.25 block and for 9 of 11 in the Mecs=1.83-2.75 block. Because a Gaussian KDE has strictly positive density everywhere, -inf can only be floating-point underflow: the bandwidth is small enough that at least one observed point, likely Swift J0243.6+6124 with vz=0_pec about 312 km/s, falls so far in the tail that the density underflows to zero. The product in Eq. 11 then makes the whole model score -inf, so the finite values (-274.54, -266.04, -188.08) merely identify which models underflow least; they do not provide a meaningful ranking. The 2D Porb-e likelihood selects a different model (sigma1=265, sigma2=30), which is then dismissed because its velocity p-value is 0, leaving no consistent statistical criterion. The likelihood also uses only the median vz=0_pec and ignores the asymmetric uncertainties in Table 2. The central claim is therefore conditional on fixing this flawed comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper compiles astrometric and radial-velocity data for 36 neutron-star high-mass X-ray binaries (NS HMXBs), derives their peculiar velocities and an estimate of the velocity at birth (vz=0_pec), and compares these with binary population synthesis (BPS) simulations. The authors test 11 natal kick models (single and bimodal Maxwellians for core-collapse and electron-capture supernovae) and, based on a 3D likelihood in the (Porb, e, vz=0_pec) space, conclude that the best-fitting model is a bimodal Maxwellian with sigma1=320 km/s for CCSNe and sigma2=80 km/s for ECSNe, with ECSN helium core masses in the range 1.83-2.25 Msun. The paper also reports a velocity-only KS comparison and discusses the role of Swift J0243.6+6124.","tokens_in":23714,"tokens_out":4238,"duration_ms":44359,"significance":"The paper addresses an important and timely question: calibrating neutron star natal kicks from a young, clean population of HMXBs. Its strengths include the enlarged sample of 36 NS HMXBs, the explicit treatment of two ECSN core-mass ranges, and the use of both dynamical and orbital information in model comparison. If the statistical comparison were robust, the derived bimodal kick prescription would be a valuable empirical input for binary evolution and gravitational-wave source studies. However, the central claim currently rests on a likelihood evaluation that is numerically unstable for most models, and the model-selection procedure is internally inconsistent. The underlying astrophysical question is worth pursuing, but the present manuscript does not yet establish the claimed best-fit parameters.","major_comments":[{"comment":"Table 3 reports log Lambda(Porb,e,vz=0_pec) = -inf for 8 of 11 models in the Mecs=1.83-2.25 block and for 9 of 11 in the Mecs=1.83-2.75 block. Since a Gaussian KDE density is strictly positive everywhere, these -inf values can only be floating-point underflow rather than genuine zero likelihoods. Consequently, the finite values (-274.54, -266.04, -188.08) merely identify which models underflow the least and do not provide a meaningful ranking. The authors need to evaluate the likelihood in log space with a numerically stable bandwidth, and to demonstrate sensitivity of the ranking to the KDE bandwidth and to the standardization of the three variables. Without this fix, the claim in Section 5 that sigma1=320, sigma2=80 is the best model is not supported by the stated evidence.","section":"Section 4.2, Table 3"},{"comment":"The likelihood product in Eq. (11) uses only the median vz=0_pec of each source and ignores the asymmetric uncertainties listed in Table 2 (for example, Swift J0243.6+6124 has vz=0_pec = 312.48+12.20/-13.16 km/s). Moreover, the text states that only 24 of the 36 sources have valid Porb and e measurements, but it does not state whether the product runs over 24 or 36 sources or how missing Porb/e values are handled. This ambiguity and the neglect of measurement uncertainties directly affect the likelihood values and the final model ranking. The analysis should marginalize over the full error distributions of all measured quantities and clearly define the sample used in Eq. (11).","section":"Section 4.2, Eq. (11)"},{"comment":"The model-selection logic is internally inconsistent. The 2D Porb-e likelihood selects the model with sigma1=265, sigma2=30 and Mecs=1.83-2.75, but this model is then discarded because its velocity KS p-value is 0. The 3D likelihood instead selects sigma1=320, sigma2=80 and Mecs=1.83-2.25. The paper does not provide a single consistent statistical criterion that distinguishes these selections; the 'best model' is therefore obtained by a post-hoc reconciliation of two incompatible rankings. A principled combined likelihood (or a joint prior and a stated model-selection rule) is needed before the abstract's claim can be accepted.","section":"Section 4.2"},{"comment":"The comparison treats vz=0_pec, derived by backward Galactic orbit integration over 10 Gyr and averaging disk-crossing velocities, as equivalent to the simulated systemic velocity vsym at the moment of NS birth. This requires that the unknown age of each HMXB does not bias the disk-crossing average, that the Galactic potential is accurate over the system's lifetime, and that unresolved binary orbital motion does not contaminate the Gaia proper motions. None of these assumptions is tested. The authors should validate the proxy by applying the same backward-integration procedure to synthetic populations with known birth velocities and ages, and propagate the resulting bias into the model comparison.","section":"Section 2.3, Section 4.1"},{"comment":"The KS tests are performed using only the median vz=0_pec values and ignore the measurement uncertainties. Given the small sample size (N=36) and the asymmetric error bars in Table 2, the p-values in Table 3 and Figures 1-2 could shift appreciably if the full error distributions were propagated through a Monte Carlo resampling scheme. In particular, the conclusion that certain models are 'eliminated' by the velocity KS test (p < 0.05) should be checked for robustness; this is not merely a cosmetic issue because the filtering is used to argue against the 2D Porb-e preferred model.","section":"Section 4.1"}],"minor_comments":[{"comment":"The text says 'the derived medium vz=0_pec values' but should say 'median'.","section":"Section 4.1"},{"comment":"Several sources have no Porb and/or e listed, while for IGR J11215-5952 and IGR J18027-2016 only eccentricity limits are given. The manuscript should explicitly state how these entries are treated in the Porb-e-vz=0_pec likelihood.","section":"Table 2 and Section 4.2"},{"comment":"The sentence 'We set up two single Maxwellian distributions' is confusing because Table 3 contains 11 models, two of which are single Maxwellians. Rephrase as 'two single-Maxwellian models'.","section":"Section 3.2"},{"comment":"In footnote 4, the phrase 'it was pointed out that vz=0_pec is not influenced by the integration time' would benefit from a precise citation or a more quantitative justification; as written it is vague.","section":"Section 2.3"},{"comment":"The side-panel density plots are too small to read, and some axis labels are ambiguous. Consider enlarging the panels and using clearer labels for the probability density.","section":"Figures 3 and 4"},{"comment":"The discussion of Swift J0243.6+6124 ('its abnormal velocity... could be inaccurate') is honest, but the claim that excluding it leaves the best model unchanged is not supported by a table or figure; please present this robustness test explicitly.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially useful paper with a new sample and an interesting astrophysical question, but the headline result is not statistically established. The -inf likelihood values in Table 3 are a red flag that the KDE-based comparison is numerically broken for most models, and the 2D/3D model-selection inconsistency further undermines the claimed best fit. I would ask the authors to redo the likelihood analysis in a numerically stable way, marginalize over measurement uncertainties, define a single model-selection criterion, and test the vz=0_pec proxy. If those changes are made, the paper could become a solid contribution to the natal kick literature. It would also be good practice to make the simulation outputs and analysis scripts available for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the headline numbers (sigma1=320, sigma2=80, Mecs 1.83-2.25) are not established by the paper's own comparison. The 3D likelihood in Section 4.2 returns -inf for most models; since a Gaussian KDE has positive density everywhere, that -inf is floating-point underflow, not a real zero. The finite values (-188 etc.) only tell you which models underflow least. So the model ranking rests on a numerical artifact. The KS test on velocities alone cannot pick one model, as the authors admit, and the 2D Porb-e test prefers a different model that is then rejected on velocity grounds. There is no consistent statistical criterion in the paper.\n\nThat said, there is real substance here. The authors expanded the Zhao et al. sample from 25 to 36 NS HMXBs with new Gaia DR3 astrometry and radial velocities, derived vz=0_pec by backward orbit integration, and ran a proper BSE population synthesis with bimodal Maxwellian kicks and two ECSN He-core mass ranges. The discussion of the high-velocity outlier Swift J0243.6+6124 is honest, and they check that the best model is unchanged when it is excluded. The references to Verbunt et al. and Igoshev et al. are fair; the resulting parameters being close to those earlier pulsar-based estimates is plausible but not evidence, because the input space was built around those values.\n\nThe soft spots are real but fixable. The likelihood needs a validated KDE bandwidth or a proper ABC/simulation-based inference, and it should propagate the asymmetric uncertainties in vz=0_pec instead of using the medians. The sample is small and not selection-corrected, so the weights and proportions (71/29) are shaky. The vz=0_pec proxy itself—disk-crossing averages from 10 Gyr backward integrations—deserves a sensitivity test, because HMXBs are younger than that but the disk-crossing average is meant to be age-independent; still, this is a reasonable assumption, not a fatal flaw.\n\nWho gets value from this? People working on the evolutionary channels of X-ray binaries and on kick prescriptions for compact-object merger population synthesis. The data table and the BPS setup are a useful resource. But I would not cite the kick parameters as calibrated values. The paper deserves a serious referee—the question is important and the sample is valuable—but it needs major revision of the statistical comparison before the headline claim is acceptable. Send it to review with someone who has both BPS and KDE experience.","headline":"Expanded HMXB sample and BPS setup are valuable, but the headline kick parameters rest on a 3D likelihood that underflows to -inf for most models, so the calibrated values aren't yet credible.","tokens_in":24331,"tokens_out":3996,"would_cite":false,"duration_ms":37410,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Neutron star natal kicks are best described by a bimodal Maxwellian with 320 km/s for core-collapse supernovae and 80 km/s for electron-capture supernovae.","keywords":["neutron star natal kicks","high-mass X-ray binaries","population synthesis","electron-capture supernovae","bimodal Maxwellian distribution","peculiar velocities","core-collapse supernovae"],"falsifier":"Take one or more HMXBs with independently known ages and orbital parameters, integrate each orbit backward for its true age rather than 10 Gyr, and compare the resulting disk-crossing velocity with the birth systemic velocity predicted by the supernova mass-loss relation; a systematic offset across several systems would show that $v^{z=0}_{\\rm pec}$ is not a faithful proxy for $v_{\\rm sym}$ and would reopen the kick fit.","tokens_in":23135,"feed_emoji":"💫","tokens_out":8047,"duration_ms":75390,"temperature":0.7,"pith_summary":"Neutron stars receive a 'natal kick' when they are born in a supernova, and the size of that kick shapes how the surviving binary moves through the Galaxy. This paper tries to measure the kick distribution by assembling 36 neutron-star high-mass X-ray binaries (HMXBs) with astrometric and radial-velocity data, computing their peculiar velocities at birth, and comparing them with a simulated population of HMXBs evolved under different kick prescriptions. The author argues that the observations favor a bimodal Maxwellian distribution: core-collapse supernovae produce neutron stars with a velocity dispersion of $\\sigma_1 = 320\\ \\mathrm{km\\,s^{-1}}$, while electron-capture supernovae produce a slower component with $\\sigma_2 = 80\\ \\mathrm{km\\,s^{-1}}$, and the helium-core mass window for the electron-capture channel is restricted to $(1.83\\!-\\!2.25)\\,M_\\odot$. If right, this gives population synthesis and binary evolution studies a concrete kick prescription and ties the slow-kick subpopulation to electron-capture supernovae.","feed_headline":"Best-fit neutron star kicks: 320 and 80 km/s","feed_subtitle":"A population synthesis match to 36 high-mass X-ray binaries picks a two-component Maxwellian kick.","key_machinery":"The load-bearing object is the kick-to-velocity relation for a binary that loses mass in a supernova, $\\vec{v}_{\\rm sym} = \\frac{M'_1}{M'_b} \\vec{v}_k - \\frac{\\Delta M_1 M_2}{M'_b M_b} \\vec{v}$, where $M_1, M'_1$ are the exploding star's pre- and post-supernova masses, $M_2$ the companion mass, $M_b, M'_b$ the total pre- and post-supernova binary masses, $\\Delta M_1$ the ejected mass, and $\\vec{v}$ the pre-supernova relative orbital velocity. This identity converts a hypothesized kick distribution into a predicted distribution of systemic velocities, so the observed space velocities can be used to rank kick models. The ranking machinery is a Monte Carlo binary population synthesis code that evolves $10^7$ binaries, followed by a Kolmogorov\\textendash{}Smirnov test on velocities and a Bayesian likelihood on the joint $P_{\\rm orb}$\\textendash{}$e$\\textendash{}$v^{z=0}_{\\rm pec}$ distribution.","core_discovery":"Using a Monte Carlo binary population synthesis model, the paper evolves $10^7$ primordial binaries and retains those that become HMXBs with a companion more massive than $8\\,M_\\odot$. For each of eleven candidate kick distributions, it compares the simulated systemic velocity $v_{\\rm sym}$ (the binary's velocity immediately after the supernova, given by a vector sum of the kick and the pre-supernova orbital velocity, scaled by mass loss) with the observationally inferred birth velocity $v^{z=0}_{\\rm pec}$ of 36 NS HMXBs. The velocity comparison alone leaves several models statistically acceptable, so the paper adds a Bayesian likelihood over the orbital period, eccentricity, and birth velocity. The model with the highest joint likelihood is a bimodal Maxwellian with $\\sigma_1 = 320\\ \\mathrm{km\\,s^{-1}}$ for core-collapse supernovae and $\\sigma_2 = 80\\ \\mathrm{km\\,s^{-1}}$ for electron-capture supernovae, with the ECSN helium-core mass in $(1.83\\!-\\!2.25)\\,M_\\odot$; the ECSN component contributes about 29% of the NS HMXBs. The paper itself notes that the supernova mechanism is the most important uncertainty and that natal kicks may depend on progenitor properties rather than a universal distribution, so the quoted parameters are the best fit within the assumed framework, not a guaranteed physical law.","pith_inferences":["If the bimodal picture is correct, low-eccentricity, long-period NS HMXBs may preferentially mark electron-capture supernova births, giving a practical way to tag the formation channel from observable orbital parameters alone.","The same comparison could be applied to Be/X-ray binaries or to black-hole HMXBs to test whether the slow-kick component is specific to ECSNe or scales with remnant mass.","Because the velocity data alone cannot separate several models, the quoted $\\sigma$ values should be read as the best available joint fit rather than a unique inversion; a larger sample with smaller astrometric errors could tighten or shift them."],"forward_implications":["Population synthesis calculations of neutron-star binaries that use a single Maxwellian kick will miss the slow ECSN component; adopting $\\sigma_1=320$, $\\sigma_2=80$ km/s with an ECSN He-core window of $(1.83\\!-\\!2.25)\\,M_\\odot$ reproduces the observed HMXB velocities, periods, and eccentricities.","Under the best model roughly 29% of NS HMXBs form through electron-capture supernovae, so the ECSN channel is a substantial, not rare, contributor to the Galactic HMXB population.","The high-velocity system Swift J0243.6+6124, with $v^{z=0}_{\\rm pec}\\approx 312$ km/s, is either an extraordinary kick or an unreliable radial-velocity measurement; the paper finds the best model survives its removal.","Kick models in which both components have dispersions below about 200 km/s are effectively excluded by the combined orbital and velocity data."],"supporting_citations":[{"why":"Supplies the binary population synthesis machinery that evolves the binaries and applies the kick prescriptions.","marker":"Hurley et al. 2002"},{"why":"Provides the rapid supernova remnant-mass recipe used for core-collapse supernova remnants.","marker":"Fryer et al. 2012"},{"why":"Gives the bimodal Maxwellian model family for young pulsar kicks that the paper compares against.","marker":"Verbunt et al. 2017"},{"why":"Supplies the single-Maxwellian $\\sigma = 265$ km/s kick baseline.","marker":"Hobbs et al. 2005"},{"why":"Provides 25 of the 36 NS HMXBs and the disk-crossing birth-velocity method used here.","marker":"Zhao et al. 2023"},{"why":"Establishes the backward orbit integration used to estimate $v^{z=0}_{\\rm pec}$.","marker":"Atri et al. 2019"},{"why":"Supplies the two ECSN helium-core mass ranges considered in the models.","marker":"Shao & Li 2018"},{"why":"Presents the bimodal kick inference for Be X-ray binaries that motivates the two-component models.","marker":"Igoshev et al. 2021"},{"why":"Provides theoretical electron-capture supernova simulations cited to justify small kicks for ECSNe.","marker":"Gessner & Janka 2018"}],"fun_headline_variants":["Bimodal Maxwellian kicks: 320 and 80 km/s for neutron stars","Neutron star natal kicks: two speeds, 320 and 80 km/s","Best-fit kicks from 36 X-ray binaries: 320 and 80 km/s","Core-collapse at 320, electron-capture at 80 km/s","Twin Maxwellian kicks explain neutron star binaries"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the observationally derived disk-crossing birth velocity of each HMXB equals the simulated systemic velocity at the moment of neutron star birth; if unresolved orbital motion, errors in the Galactic potential, or the unknown age of each system biases that proxy, the model ranking and the claimed kick parameters change.","fun_headline_variants_meta":{"raw":{"variants":["Bimodal Maxwellian kicks: 320 and 80 km/s for neutron stars","Neutron star natal kicks: two speeds, 320 and 80 km/s","Best-fit kicks from 36 X-ray binaries: 320 and 80 km/s","Core-collapse at 320, electron-capture at 80 km/s","Twin Maxwellian kicks explain neutron star binaries"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3331,"prompt_tokens":1082,"completion_tokens":2249,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":698,"completion_tokens_details":{"reasoning_tokens":2148}},"tokens_in":698,"tokens_out":2249,"duration_ms":17187,"temperature":1.0,"reasoning_tokens":2148,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:46:16.136761+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one or more HMXBs with independently known ages and orbital parameters, integrate each orbit backward for its true age rather than 10 Gyr, and compare the resulting disk-crossing velocity with the birth systemic velocity predicted by the supernova mass-loss relation; a systematic offset across several systems would show that $v^{z=0}_{\\rm pec}$ is not a faithful proxy for $v_{\\rm sym}$ and would reopen the kick fit.","supporting_citations":[],"review_version":1}